Skip to contentSkip to conversation

Remove use of kmap()

PATCH Applied [email protected] 4 participants 22 messages Jun 22, 2021 – Jul 15, 2021

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

cover Remove use of kmap()

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS protection is introduced.[1]

These kmap() usages don't need to be global and work fine as thread local mappings.

Replace these kmap() calls with kmap_local_page() which is more appropriate.

The only final use of kmap() in the RDMA subsystem is in the qib driver which is pretty old at this point. The use is pretty convoluted and I doubt systems using that driver are using persistent memory. So it is left as is. If this is a problem I can dig into converting it as well.

[1] https://lore.kernel.org/lkml/[email protected]/

Ira Weiny (4):
RDMA/hfi1: Remove use of kmap()
RDMA/i40iw: Remove use of kmap()
RDMA/siw: Remove kmap()
RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

 drivers/infiniband/hw/hfi1/sdma.c      |  4 +--
 drivers/infiniband/hw/i40iw/i40iw_cm.c | 10 +++---
 drivers/infiniband/sw/siw/siw_qp_tx.c  | 47 +++++++++++++++-----------
 3 files changed, 34 insertions(+), 27 deletions(-)
Signature · 1 line
2.28.0.rc0.12.gb6a658bd00c9
#1 lore raw

PATCH 1/4 RDMA/hfi1: Remove use of kmap()

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS protection is introduced.[1]

The kmap() used in sdma does not need to be global. Use the new kmap_local_page() which works with PKS and may provide better performance for this thread local mapping anyway.

[1] https://lore.kernel.org/lkml/[email protected]/

---
 drivers/infiniband/hw/hfi1/sdma.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/infiniband/hw/hfi1/sdma.c +2 -2
index 46b5290b2839..af43dcbb0928 100644
@@ -3130,7 +3130,7 @@ int ext_coal_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx,
 		}
 
 		if (type == SDMA_MAP_PAGE) {
-			kvaddr = kmap(page);
+			kvaddr = kmap_local_page(page);
 			kvaddr += offset;
 		} else if (WARN_ON(!kvaddr)) {
 			__sdma_txclean(dd, tx);
@@ -3140,7 +3140,7 @@ int ext_coal_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx,
 		memcpy(tx->coalesce_buf + tx->coalesce_idx, kvaddr, len);
 		tx->coalesce_idx += len;
 		if (type == SDMA_MAP_PAGE)
-			kunmap(page);
+			kunmap_local(kvaddr);
 
 		/* If there is more data, return */
 		if (tx->tlen - tx->coalesce_idx)
Signature · 1 line
2.28.0.rc0.12.gb6a658bd00c9
maintainer Applied
#2 lore raw

Re: [PATCH 1/4] RDMA/hfi1: Remove use of kmap()

Quoted 9 lines · On Mon, Jun 21, 2021 at 11:14:19PM -0700, [email protected] wrote:

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS protection is introduced.[1]

The kmap() used in sdma does not need to be global. Use the new kmap_local_page() which works with PKS and may provide better performance for this thread local mapping anyway.

Quoted 6 lines

[1] https://lore.kernel.org/lkml/[email protected]/

---
 drivers/infiniband/hw/hfi1/sdma.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

Applied to for-next, thanks

Jason

#3 lore raw

PATCH 2/4 RDMA/i40iw: Remove use of kmap()

In reply to [email protected]

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS protection is introduced.[1]

The kmap() used in the i40iw CM driver is thread local. Therefore kmap_local_page() sufficient to use and may provide performance benefits as well. kmap_local_page() will work with device dax and pgmap protected pages.

Use kmap_local_page() instead of kmap().

[1] https://lore.kernel.org/lkml/[email protected]/

---
 drivers/infiniband/hw/i40iw/i40iw_cm.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)
drivers/infiniband/hw/i40iw/i40iw_cm.c +5 -5
index ac65c8237b2e..13e40ebd5322 100644
@@ -3725,7 +3725,7 @@ int i40iw_accept(struct iw_cm_id *cm_id, struct iw_cm_conn_param *conn_param)
 		ibmr->device = iwpd->ibpd.device;
 		iwqp->lsmm_mr = ibmr;
 		if (iwqp->page)
-			iwqp->sc_qp.qp_uk.sq_base = kmap(iwqp->page);
+			iwqp->sc_qp.qp_uk.sq_base = kmap_local_page(iwqp->page);
 		dev->iw_priv_qp_ops->qp_send_lsmm(&iwqp->sc_qp,
 							iwqp->ietf_mem.va,
 							(accept.size + conn_param->private_data_len),
@@ -3733,12 +3733,12 @@ int i40iw_accept(struct iw_cm_id *cm_id, struct iw_cm_conn_param *conn_param)
 
 	} else {
 		if (iwqp->page)
-			iwqp->sc_qp.qp_uk.sq_base = kmap(iwqp->page);
+			iwqp->sc_qp.qp_uk.sq_base = kmap_local_page(iwqp->page);
 		dev->iw_priv_qp_ops->qp_send_lsmm(&iwqp->sc_qp, NULL, 0, 0);
 	}
 
 	if (iwqp->page)
-		kunmap(iwqp->page);
+		kunmap_local(iwqp->sc_qp.qp_uk.sq_base);
 
 	iwqp->cm_id = cm_id;
 	cm_node->cm_id = cm_id;
@@ -4106,10 +4106,10 @@ static void i40iw_cm_event_connected(struct i40iw_cm_event *event)
 	i40iw_cm_init_tsa_conn(iwqp, cm_node);
 	read0 = (cm_node->send_rdma0_op == SEND_RDMA_READ_ZERO);
 	if (iwqp->page)
-		iwqp->sc_qp.qp_uk.sq_base = kmap(iwqp->page);
+		iwqp->sc_qp.qp_uk.sq_base = kmap_local_page(iwqp->page);
 	dev->iw_priv_qp_ops->qp_send_rtt(&iwqp->sc_qp, read0);
 	if (iwqp->page)
-		kunmap(iwqp->page);
+		kunmap_local(iwqp->sc_qp.qp_uk.sq_base);
 
 	memset(&attr, 0, sizeof(attr));
 	attr.qp_state = IB_QPS_RTS;
Signature · 1 line
2.28.0.rc0.12.gb6a658bd00c9
#4 lore raw

Re: [PATCH 2/4] RDMA/i40iw: Remove use of kmap()

Quoted 10 lines · On Mon, Jun 21, 2021 at 11:14:20PM -0700, [email protected] wrote:

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS protection is introduced.[1]

The kmap() used in the i40iw CM driver is thread local. Therefore kmap_local_page() sufficient to use and may provide performance benefits as well. kmap_local_page() will work with device dax and pgmap protected pages.

Quoted 8 lines

Use kmap_local_page() instead of kmap().

[1] https://lore.kernel.org/lkml/[email protected]/

---
 drivers/infiniband/hw/i40iw/i40iw_cm.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

This needs to be resent against irdma instead

Jason

#5 lore raw

PATCH v2 1/1 RDMA/irdma: Remove use of kmap()

In reply to [email protected]

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS protection is introduced.[1]

The kmap() used in the irdma CM driver is thread local. Therefore kmap_local_page() is sufficient to use and may provide performance benefits as well. kmap_local_page() will work with device dax and pgmap protected pages.

Use kmap_local_page() instead of kmap().

[1] https://lore.kernel.org/lkml/[email protected]/

---
Changes for V2:
Move to the new irdma driver for 5.14

---
 drivers/infiniband/hw/irdma/cm.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
drivers/infiniband/hw/irdma/cm.c +4 -4
index 3d2bdb033a54..6b62299abfbb 100644
@@ -3675,14 +3675,14 @@ int irdma_accept(struct iw_cm_id *cm_id, struct iw_cm_conn_param *conn_param)
 	ibmr->device = iwpd->ibpd.device;
 	iwqp->lsmm_mr = ibmr;
 	if (iwqp->page)
-		iwqp->sc_qp.qp_uk.sq_base = kmap(iwqp->page);
+		iwqp->sc_qp.qp_uk.sq_base = kmap_local_page(iwqp->page);
 
 	cm_node->lsmm_size = accept.size + conn_param->private_data_len;
 	irdma_sc_send_lsmm(&iwqp->sc_qp, iwqp->ietf_mem.va, cm_node->lsmm_size,
 			   ibmr->lkey);
 
 	if (iwqp->page)
-		kunmap(iwqp->page);
+		kunmap_local(iwqp->sc_qp.qp_uk.sq_base);
 
 	iwqp->cm_id = cm_id;
 	cm_node->cm_id = cm_id;
@@ -4093,10 +4093,10 @@ static void irdma_cm_event_connected(struct irdma_cm_event *event)
 	irdma_cm_init_tsa_conn(iwqp, cm_node);
 	read0 = (cm_node->send_rdma0_op == SEND_RDMA_READ_ZERO);
 	if (iwqp->page)
-		iwqp->sc_qp.qp_uk.sq_base = kmap(iwqp->page);
+		iwqp->sc_qp.qp_uk.sq_base = kmap_local_page(iwqp->page);
 	irdma_sc_send_rtt(&iwqp->sc_qp, read0);
 	if (iwqp->page)
-		kunmap(iwqp->page);
+		kunmap_local(iwqp->sc_qp.qp_uk.sq_base);
 
 	attr.qp_state = IB_QPS_RTS;
 	cm_node->qhash_set = false;
Signature · 1 line
2.28.0.rc0.12.gb6a658bd00c9
maintainer Applied
#6 lore raw

Re: [PATCH V2] RDMA/irdma: Remove use of kmap()

Quoted 14 lines · On Tue, Jun 22, 2021 at 09:56:22AM -0700, [email protected] wrote:

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS protection is introduced.[1]

The kmap() used in the irdma CM driver is thread local. Therefore kmap_local_page() is sufficient to use and may provide performance benefits as well. kmap_local_page() will work with device dax and pgmap protected pages.

Use kmap_local_page() instead of kmap().

[1] https://lore.kernel.org/lkml/[email protected]/

Quoted 7 lines

---
Changes for V2:
Move to the new irdma driver for 5.14

---
 drivers/infiniband/hw/irdma/cm.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

Applied to for-next, thanks

Jason

#7 lore raw

PATCH 3/4 RDMA/siw: Remove kmap()

In reply to [email protected]

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS protection is introduced.[1]

These uses of kmap() in the SIW driver are thread local. Therefore kmap_local_page() is sufficient to use and will work with pgmap protected pages when those are implemnted.

There is one more use of kmap() in this driver which is split into its own patch because kmap_local_page() has strict ordering rules and the use of the kmap_mask over multiple segments must be handled carefully. Therefore, that conversion is handled in a stand alone patch.

Use kmap_local_page() instead of kmap() in the 'easy' cases.

[1] https://lore.kernel.org/lkml/[email protected]/

---
 drivers/infiniband/sw/siw/siw_qp_tx.c | 14 ++++++++------
 1 file changed, 8 insertions(+), 6 deletions(-)
drivers/infiniband/sw/siw/siw_qp_tx.c +8 -6
index 7989c4043db4..db68a10d12cd 100644
@@ -76,7 +76,7 @@ static int siw_try_1seg(struct siw_iwarp_tx *c_tx, void *paddr)
 			if (unlikely(!p))
 				return -EFAULT;
 
-			buffer = kmap(p);
+			buffer = kmap_local_page(p);
 
 			if (likely(PAGE_SIZE - off >= bytes)) {
 				memcpy(paddr, buffer + off, bytes);
@@ -84,7 +84,7 @@ static int siw_try_1seg(struct siw_iwarp_tx *c_tx, void *paddr)
 				unsigned long part = bytes - (PAGE_SIZE - off);
 
 				memcpy(paddr, buffer + off, part);
-				kunmap(p);
+				kunmap_local(buffer);
 
 				if (!mem->is_pbl)
 					p = siw_get_upage(mem->umem,
@@ -96,10 +96,10 @@ static int siw_try_1seg(struct siw_iwarp_tx *c_tx, void *paddr)
 				if (unlikely(!p))
 					return -EFAULT;
 
-				buffer = kmap(p);
+				buffer = kmap_local_page(p);
 				memcpy(paddr + part, buffer, bytes - part);
 			}
-			kunmap(p);
+			kunmap_local(buffer);
 		}
 	}
 	return (int)bytes;
@@ -485,6 +485,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 
 		while (sge_len) {
 			size_t plen = min((int)PAGE_SIZE - fp_off, sge_len);
+			void *kaddr;
 
 			if (!is_kva) {
 				struct page *p;
@@ -517,10 +518,11 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 							iov[seg].iov_base,
 							plen);
 				} else if (do_crc) {
+					kaddr = kmap_local_page(p);
 					crypto_shash_update(c_tx->mpa_crc_hd,
-							    kmap(p) + fp_off,
+							    kaddr + fp_off,
 							    plen);
-					kunmap(p);
+					kunmap_local(kaddr);
 				}
 			} else {
 				u64 va = sge->laddr + sge_off;
Signature · 1 line
2.28.0.rc0.12.gb6a658bd00c9
maintainer Applied
#8 lore raw

Re: [PATCH 3/4] RDMA/siw: Remove kmap()

Quoted 14 lines · On Mon, Jun 21, 2021 at 11:14:21PM -0700, [email protected] wrote:

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS protection is introduced.[1]

These uses of kmap() in the SIW driver are thread local. Therefore kmap_local_page() is sufficient to use and will work with pgmap protected pages when those are implemnted.

There is one more use of kmap() in this driver which is split into its own patch because kmap_local_page() has strict ordering rules and the use of the kmap_mask over multiple segments must be handled carefully. Therefore, that conversion is handled in a stand alone patch.

Quoted 8 lines

Use kmap_local_page() instead of kmap() in the 'easy' cases.

[1] https://lore.kernel.org/lkml/[email protected]/

---
 drivers/infiniband/sw/siw/siw_qp_tx.c | 14 ++++++++------
 1 file changed, 8 insertions(+), 6 deletions(-)

Applied to for-next, thanks

Jason

#9 lore raw

PATCH 4/4 RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

In reply to [email protected]

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS protection is introduced.[1]

The use of kmap() in siw_tx_hdt() is all thread local therefore kmap_local_page() is a sufficient replacement and will work with pgmap protected pages when those are implemented.

kmap_local_page() mappings are tracked in a stack and must be unmapped in the opposite order they were mapped in.

siw_tx_hdt() tracks pages used in a page_array. It uses that array to unmap pages which were mapped on function exit. Not all entries in the array are mapped and this is tracked in kmap_mask.

kunmap_local() takes a mapped address rather than a page. Declare a mapped address array, page_array_addr, of the same size as the page array to be used for unmapping.

Use kmap_local_page() instead of kmap() to map pages in the page_array.

Because segments are mapped into the page array in increasing index order, modify siw_unmap_pages() to unmap pages in decreasing order.

The kmap_mask is no longer needed as the lack of an address in the address array can indicate no unmap is required.

[1] https://lore.kernel.org/lkml/[email protected]/

---
 drivers/infiniband/sw/siw/siw_qp_tx.c | 35 +++++++++++++++------------
 1 file changed, 20 insertions(+), 15 deletions(-)
drivers/infiniband/sw/siw/siw_qp_tx.c +20 -15
index db68a10d12cd..e70aba23f6e7 100644
@@ -396,13 +396,17 @@ static int siw_0copy_tx(struct socket *s, struct page **page,
 
 #define MAX_TRAILER (MPA_CRC_SIZE + 4)
 
-static void siw_unmap_pages(struct page **pp, unsigned long kmap_mask)
+static void siw_unmap_pages(void **addrs, int len)
 {
-	while (kmap_mask) {
-		if (kmap_mask & BIT(0))
-			kunmap(*pp);
-		pp++;
-		kmap_mask >>= 1;
+	int i;
+
+	/*
+	 * Work backwards through the array to honor the kmap_local_page()
+	 * ordering requirements.
+	 */
+	for (i = (len-1); i >= 0; i--) {
+		if (addrs[i])
+			kunmap_local(addrs[i]);
 	}
 }
 
@@ -427,13 +431,15 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 	struct siw_sge *sge = &wqe->sqe.sge[c_tx->sge_idx];
 	struct kvec iov[MAX_ARRAY];
 	struct page *page_array[MAX_ARRAY];
+	void *page_array_addr[MAX_ARRAY];
 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR };
 
 	int seg = 0, do_crc = c_tx->do_crc, is_kva = 0, rv;
 	unsigned int data_len = c_tx->bytes_unsent, hdr_len = 0, trl_len = 0,
 		     sge_off = c_tx->sge_off, sge_idx = c_tx->sge_idx,
 		     pbl_idx = c_tx->pbl_idx;
-	unsigned long kmap_mask = 0L;
+
+	memset(page_array_addr, 0, sizeof(page_array_addr));
 
 	if (c_tx->state == SIW_SEND_HDR) {
 		if (c_tx->use_sendpage) {
@@ -498,7 +504,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 					p = siw_get_upage(mem->umem,
 							  sge->laddr + sge_off);
 				if (unlikely(!p)) {
-					siw_unmap_pages(page_array, kmap_mask);
+					siw_unmap_pages(page_array_addr, MAX_ARRAY);
 					wqe->processed -= c_tx->bytes_unsent;
 					rv = -EFAULT;
 					goto done_crc;
@@ -506,11 +512,10 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 				page_array[seg] = p;
 
 				if (!c_tx->use_sendpage) {
-					iov[seg].iov_base = kmap(p) + fp_off;
-					iov[seg].iov_len = plen;
+					page_array_addr[seg] = kmap_local_page(page_array[seg]);
 
-					/* Remember for later kunmap() */
-					kmap_mask |= BIT(seg);
+					iov[seg].iov_base = page_array_addr[seg] + fp_off;
+					iov[seg].iov_len = plen;
 
 					if (do_crc)
 						crypto_shash_update(
@@ -518,7 +523,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 							iov[seg].iov_base,
 							plen);
 				} else if (do_crc) {
-					kaddr = kmap_local_page(p);
+					kaddr = kmap_local_page(page_array[seg]);
 					crypto_shash_update(c_tx->mpa_crc_hd,
 							    kaddr + fp_off,
 							    plen);
@@ -542,7 +547,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 
 			if (++seg > (int)MAX_ARRAY) {
 				siw_dbg_qp(tx_qp(c_tx), "to many fragments\n");
-				siw_unmap_pages(page_array, kmap_mask);
+				siw_unmap_pages(page_array_addr, MAX_ARRAY);
 				wqe->processed -= c_tx->bytes_unsent;
 				rv = -EMSGSIZE;
 				goto done_crc;
@@ -593,7 +598,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 	} else {
 		rv = kernel_sendmsg(s, &msg, iov, seg + 1,
 				    hdr_len + data_len + trl_len);
-		siw_unmap_pages(page_array, kmap_mask);
+		siw_unmap_pages(page_array_addr, MAX_ARRAY);
 	}
 	if (rv < (int)hdr_len) {
 		/* Not even complete hdr pushed or negative rv */
Signature · 1 line
2.28.0.rc0.12.gb6a658bd00c9
#10 lore raw

Re: [PATCH 4/4] RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

[email protected] wrote: -----

Quoted 34 lines

To: "Jason Gunthorpe" <[email protected]>
From: [email protected]
Date: 06/22/2021 08:14AM
Cc: "Ira Weiny" <[email protected]>, "Mike Marciniszyn" <[email protected]>, "Dennis Dalessandro" <[email protected]>, "Doug Ledford" <[email protected]>, "Faisal Latif" <[email protected]>, "Shiraz Saleem" <[email protected]>, "Bernard Metzler" <[email protected]>, "Kamal Heib" <[email protected]>, [email protected], [email protected]
Subject: [EXTERNAL] [PATCH 4/4] RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS
protection is introduced.[1]

The use of kmap() in siw_tx_hdt() is all thread local therefore kmap_local_page() is a sufficient replacement and will work with pgmap
protected pages when those are implemented.

kmap_local_page() mappings are tracked in a stack and must be unmapped
in the opposite order they were mapped in.

siw_tx_hdt() tracks pages used in a page_array. It uses that array to
unmap pages which were mapped on function exit. Not all entries in the
array are mapped and this is tracked in kmap_mask.

Quoted 4 lines

kunmap_local() takes a mapped address rather than a page. Declare a mapped address array, page_array_addr, of the same size as the page array to be used for unmapping.

Hi Ira, thanks for taking care of that!

I think we can avoid introducing another 'page_array_addr[]' array here, which must be zeroed first and completely searched for valid mappings during unmap, and also further bloats the stack size of siw_tx_hdt(). I think we can go away with the already available iov[].iov_base addresses array, masking addresses with PAGE_MASK during unmapping to mask any first byte offset. All kmap_local_page() mapping end up at that list. For unmapping we can still rely on the kmap_mask bit field, which is more efficient to initialize and search for valid mappings. Ordering during unmapping can be guaranteed if we parse the bitmask in reverse order. Let me know if you prefer me to propose a change -- that siw_tx_hdt() thing became rather complex I have to admit!

Best,
Bernard.

Quoted 132 lines

Use kmap_local_page() instead of kmap() to map pages in the page_array.

Because segments are mapped into the page array in increasing index order, modify siw_unmap_pages() to unmap pages in decreasing order.

The kmap_mask is no longer needed as the lack of an address in the address array can indicate no unmap is required.

[1]
INVALID URI REMOVED lkml_20201009195033.3208459-2D59-2Dira.weiny-40intel.com_&d=DwIDAg&c= jf_iaSHvJObTbx-siA1ZOg&r=2TaYXQ0T-r8ZO1PP1alNwU_QJcRRLfmYTAgd3QCvqSc& m=wnRcc-qyXV_X7kyQfFYL6XPgmmakQxmo44BmjIon-w0&s=Y0aiKJ4EHZY8FJlI-uiPr xcBE95kmgn3iEz3p8d5VF4&e=

---
drivers/infiniband/sw/siw/siw_qp_tx.c | 35
+++++++++++++++------------

1 file changed, 20 insertions(+), 15 deletions(-)
drivers/infiniband/sw/siw/siw_qp_tx.c +10 -6
index db68a10d12cd..e70aba23f6e7 100644
@@ -396,13 +396,17 @@ static int siw_0copy_tx(struct socket *s,
 struct page **page,
 
 #define MAX_TRAILER (MPA_CRC_SIZE + 4)
 
-static void siw_unmap_pages(struct page **pp, unsigned long
 kmap_mask)
+static void siw_unmap_pages(void **addrs, int len)
 {
-	while (kmap_mask) {
-		if (kmap_mask & BIT(0))
-			kunmap(*pp);
-		pp++;
-		kmap_mask >>= 1;
+	int i;
+
+	/*
+	 * Work backwards through the array to honor the kmap_local_page()
+	 * ordering requirements.
+	 */
+	for (i = (len-1); i >= 0; i--) {
+		if (addrs[i])
+			kunmap_local(addrs[i]);
 	}

}

drivers/infiniband/sw/siw/siw_qp_tx.c +3 -1
@@ -427,13 +431,15 @@ static int siw_tx_hdt(struct siw_iwarp_tx
 *c_tx, struct socket *s)
 	struct siw_sge *sge = &wqe->sqe.sge[c_tx->sge_idx];
 	struct kvec iov[MAX_ARRAY];
 	struct page *page_array[MAX_ARRAY];
+	void *page_array_addr[MAX_ARRAY];
 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR };
 
 	int seg = 0, do_crc = c_tx->do_crc, is_kva = 0, rv;
 	unsigned int data_len = c_tx->bytes_unsent, hdr_len = 0, trl_len =
 0,
 		     sge_off = c_tx->sge_off, sge_idx = c_tx->sge_idx,
 		     pbl_idx = c_tx->pbl_idx;
-	unsigned long kmap_mask = 0L;
+
+	memset(page_array_addr, 0, sizeof(page_array_addr));
 
	if (c_tx->state == SIW_SEND_HDR) {
		if (c_tx->use_sendpage) {
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -498,7 +504,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 					p = siw_get_upage(mem->umem,
 							  sge->laddr + sge_off);
 				if (unlikely(!p)) {
-					siw_unmap_pages(page_array, kmap_mask);
+					siw_unmap_pages(page_array_addr, MAX_ARRAY);
 					wqe->processed -= c_tx->bytes_unsent;
 					rv = -EFAULT;
					goto done_crc;
drivers/infiniband/sw/siw/siw_qp_tx.c +3 -4
@@ -506,11 +512,10 @@ static int siw_tx_hdt(struct siw_iwarp_tx
 *c_tx, struct socket *s)
 				page_array[seg] = p;
 
 				if (!c_tx->use_sendpage) {
-					iov[seg].iov_base = kmap(p) + fp_off;
-					iov[seg].iov_len = plen;
+					page_array_addr[seg] = kmap_local_page(page_array[seg]);
 
-					/* Remember for later kunmap() */
-					kmap_mask |= BIT(seg);
+					iov[seg].iov_base = page_array_addr[seg] + fp_off;
+					iov[seg].iov_len = plen;
 
 					if (do_crc)
						crypto_shash_update(
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -518,7 +523,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 							iov[seg].iov_base,
 							plen);
 				} else if (do_crc) {
-					kaddr = kmap_local_page(p);
+					kaddr = kmap_local_page(page_array[seg]);
 					crypto_shash_update(c_tx->mpa_crc_hd,
 							    kaddr + fp_off,
							    plen);
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -542,7 +547,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 
 			if (++seg > (int)MAX_ARRAY) {
 				siw_dbg_qp(tx_qp(c_tx), "to many fragments\n");
-				siw_unmap_pages(page_array, kmap_mask);
+				siw_unmap_pages(page_array_addr, MAX_ARRAY);
 				wqe->processed -= c_tx->bytes_unsent;
 				rv = -EMSGSIZE;
				goto done_crc;
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -593,7 +598,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 	} else {
 		rv = kernel_sendmsg(s, &msg, iov, seg + 1,
 				    hdr_len + data_len + trl_len);
-		siw_unmap_pages(page_array, kmap_mask);
+		siw_unmap_pages(page_array_addr, MAX_ARRAY);
 	}
 	if (rv < (int)hdr_len) {
		/* Not even complete hdr pushed or negative rv */
Signature · 1 line
2.28.0.rc0.12.gb6a658bd00c9
author
#11 lore raw

Re: [PATCH 4/4] RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

Quoted 43 lines · Bernard Metzler

[email protected] wrote: -----

Quoted 34 lines

To: "Jason Gunthorpe" <[email protected]>
From: [email protected]
Date: 06/22/2021 08:14AM
Cc: "Ira Weiny" <[email protected]>, "Mike Marciniszyn" <[email protected]>, "Dennis Dalessandro" <[email protected]>, "Doug Ledford" <[email protected]>, "Faisal Latif" <[email protected]>, "Shiraz Saleem" <[email protected]>, "Bernard Metzler" <[email protected]>, "Kamal Heib" <[email protected]>, [email protected], [email protected]
Subject: [EXTERNAL] [PATCH 4/4] RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS
protection is introduced.[1]

The use of kmap() in siw_tx_hdt() is all thread local therefore kmap_local_page() is a sufficient replacement and will work with pgmap
protected pages when those are implemented.

kmap_local_page() mappings are tracked in a stack and must be unmapped
in the opposite order they were mapped in.

siw_tx_hdt() tracks pages used in a page_array. It uses that array to
unmap pages which were mapped on function exit. Not all entries in the
array are mapped and this is tracked in kmap_mask.

Quoted 4 lines

kunmap_local() takes a mapped address rather than a page. Declare a mapped address array, page_array_addr, of the same size as the page array to be used for unmapping.

Hi Ira, thanks for taking care of that!

Quoted 13 lines

I think we can avoid introducing another 'page_array_addr[]' array here, which must be zeroed first and completely searched for valid mappings during unmap, and also further bloats the stack size of siw_tx_hdt(). I think we can go away with the already available iov[].iov_base addresses array, masking addresses with PAGE_MASK during unmapping to mask any first byte offset. All kmap_local_page() mapping end up at that list. For unmapping we can still rely on the kmap_mask bit field, which is more efficient to initialize and search for valid mappings. Ordering during unmapping can be guaranteed if we parse the bitmask in reverse order. Let me know if you prefer me to propose a change -- that siw_tx_hdt() thing became rather complex I have to admit!

Seems not too bad, V2 sent.

I was concerned with the additional stack size but only 28 pointers (If I did my math right) did not seem too bad. It is redundant though so lets see if I've gotten V2 right.

Thanks!
Ira

Quoted 137 lines

Best,
Bernard.

Quoted 132 lines

Use kmap_local_page() instead of kmap() to map pages in the page_array.

Because segments are mapped into the page array in increasing index order, modify siw_unmap_pages() to unmap pages in decreasing order.

The kmap_mask is no longer needed as the lack of an address in the address array can indicate no unmap is required.

[1]
INVALID URI REMOVED lkml_20201009195033.3208459-2D59-2Dira.weiny-40intel.com_&d=DwIDAg&c= jf_iaSHvJObTbx-siA1ZOg&r=2TaYXQ0T-r8ZO1PP1alNwU_QJcRRLfmYTAgd3QCvqSc& m=wnRcc-qyXV_X7kyQfFYL6XPgmmakQxmo44BmjIon-w0&s=Y0aiKJ4EHZY8FJlI-uiPr xcBE95kmgn3iEz3p8d5VF4&e=

---
drivers/infiniband/sw/siw/siw_qp_tx.c | 35
+++++++++++++++------------

1 file changed, 20 insertions(+), 15 deletions(-)
drivers/infiniband/sw/siw/siw_qp_tx.c +10 -6
index db68a10d12cd..e70aba23f6e7 100644
@@ -396,13 +396,17 @@ static int siw_0copy_tx(struct socket *s,
 struct page **page,
 
 #define MAX_TRAILER (MPA_CRC_SIZE + 4)
 
-static void siw_unmap_pages(struct page **pp, unsigned long
 kmap_mask)
+static void siw_unmap_pages(void **addrs, int len)
 {
-	while (kmap_mask) {
-		if (kmap_mask & BIT(0))
-			kunmap(*pp);
-		pp++;
-		kmap_mask >>= 1;
+	int i;
+
+	/*
+	 * Work backwards through the array to honor the kmap_local_page()
+	 * ordering requirements.
+	 */
+	for (i = (len-1); i >= 0; i--) {
+		if (addrs[i])
+			kunmap_local(addrs[i]);
 	}

}

drivers/infiniband/sw/siw/siw_qp_tx.c +3 -1
@@ -427,13 +431,15 @@ static int siw_tx_hdt(struct siw_iwarp_tx
 *c_tx, struct socket *s)
 	struct siw_sge *sge = &wqe->sqe.sge[c_tx->sge_idx];
 	struct kvec iov[MAX_ARRAY];
 	struct page *page_array[MAX_ARRAY];
+	void *page_array_addr[MAX_ARRAY];
 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR };
 
 	int seg = 0, do_crc = c_tx->do_crc, is_kva = 0, rv;
 	unsigned int data_len = c_tx->bytes_unsent, hdr_len = 0, trl_len =
 0,
 		     sge_off = c_tx->sge_off, sge_idx = c_tx->sge_idx,
 		     pbl_idx = c_tx->pbl_idx;
-	unsigned long kmap_mask = 0L;
+
+	memset(page_array_addr, 0, sizeof(page_array_addr));
 
	if (c_tx->state == SIW_SEND_HDR) {
		if (c_tx->use_sendpage) {
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -498,7 +504,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 					p = siw_get_upage(mem->umem,
 							  sge->laddr + sge_off);
 				if (unlikely(!p)) {
-					siw_unmap_pages(page_array, kmap_mask);
+					siw_unmap_pages(page_array_addr, MAX_ARRAY);
 					wqe->processed -= c_tx->bytes_unsent;
 					rv = -EFAULT;
					goto done_crc;
drivers/infiniband/sw/siw/siw_qp_tx.c +3 -4
@@ -506,11 +512,10 @@ static int siw_tx_hdt(struct siw_iwarp_tx
 *c_tx, struct socket *s)
 				page_array[seg] = p;
 
 				if (!c_tx->use_sendpage) {
-					iov[seg].iov_base = kmap(p) + fp_off;
-					iov[seg].iov_len = plen;
+					page_array_addr[seg] = kmap_local_page(page_array[seg]);
 
-					/* Remember for later kunmap() */
-					kmap_mask |= BIT(seg);
+					iov[seg].iov_base = page_array_addr[seg] + fp_off;
+					iov[seg].iov_len = plen;
 
 					if (do_crc)
						crypto_shash_update(
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -518,7 +523,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 							iov[seg].iov_base,
 							plen);
 				} else if (do_crc) {
-					kaddr = kmap_local_page(p);
+					kaddr = kmap_local_page(page_array[seg]);
 					crypto_shash_update(c_tx->mpa_crc_hd,
 							    kaddr + fp_off,
							    plen);
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -542,7 +547,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 
 			if (++seg > (int)MAX_ARRAY) {
 				siw_dbg_qp(tx_qp(c_tx), "to many fragments\n");
-				siw_unmap_pages(page_array, kmap_mask);
+				siw_unmap_pages(page_array_addr, MAX_ARRAY);
 				wqe->processed -= c_tx->bytes_unsent;
 				rv = -EMSGSIZE;
				goto done_crc;
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -593,7 +598,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 	} else {
 		rv = kernel_sendmsg(s, &msg, iov, seg + 1,
 				    hdr_len + data_len + trl_len);
-		siw_unmap_pages(page_array, kmap_mask);
+		siw_unmap_pages(page_array_addr, MAX_ARRAY);
 	}
 	if (rv < (int)hdr_len) {
		/* Not even complete hdr pushed or negative rv */
Signature · 1 line
2.28.0.rc0.12.gb6a658bd00c9
#12 lore raw

PATCH v2 1/1 RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

In reply to [email protected]

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS protection is introduced.[1]

The use of kmap() in siw_tx_hdt() is all thread local therefore kmap_local_page() is a sufficient replacement and will work with pgmap protected pages when those are implemented.

siw_tx_hdt() tracks pages used in a page_array. It uses that array to unmap pages which were mapped on function exit. Not all entries in the array are mapped and this is tracked in kmap_mask.

kunmap_local() takes a mapped address rather than a page. Alter siw_unmap_pages() to take the iov array to reuse the iov_base address of each mapping. Use PAGE_MASK to get the proper address for kunmap_local().

kmap_local_page() mappings are tracked in a stack and must be unmapped in the opposite order they were mapped in. Because segments are mapped into the page array in increasing index order, modify siw_unmap_pages() to unmap pages in decreasing order.

Use kmap_local_page() instead of kmap() to map pages in the page_array.

[1] https://lore.kernel.org/lkml/[email protected]/

---
Changes for V2:

	From Bernard
		Reuse iov[].iov_base rather than declaring another array of
		pointers and preserve the use of kmap_mask to know which iov's
		were kmapped.
---
 drivers/infiniband/sw/siw/siw_qp_tx.c | 32 +++++++++++++++++----------
 1 file changed, 20 insertions(+), 12 deletions(-)
drivers/infiniband/sw/siw/siw_qp_tx.c +20 -12
index db68a10d12cd..fd3b9e6a67d7 100644
@@ -396,13 +396,20 @@ static int siw_0copy_tx(struct socket *s, struct page **page,
 
 #define MAX_TRAILER (MPA_CRC_SIZE + 4)
 
-static void siw_unmap_pages(struct page **pp, unsigned long kmap_mask)
+static void siw_unmap_pages(struct kvec *iov, unsigned long kmap_mask, int len)
 {
-	while (kmap_mask) {
-		if (kmap_mask & BIT(0))
-			kunmap(*pp);
-		pp++;
-		kmap_mask >>= 1;
+	int i;
+
+	/*
+	 * Work backwards through the array to honor the kmap_local_page()
+	 * ordering requirements.
+	 */
+	for (i = (len-1); i >= 0; i--) {
+		if (kmap_mask & BIT(i)) {
+			unsigned long addr = (unsigned long)iov[i].iov_base;
+
+			kunmap_local((void *)(addr & PAGE_MASK));
+		}
 	}
 }
 
@@ -498,7 +505,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 					p = siw_get_upage(mem->umem,
 							  sge->laddr + sge_off);
 				if (unlikely(!p)) {
-					siw_unmap_pages(page_array, kmap_mask);
+					siw_unmap_pages(iov, kmap_mask, MAX_ARRAY);
 					wqe->processed -= c_tx->bytes_unsent;
 					rv = -EFAULT;
 					goto done_crc;
@@ -506,11 +513,12 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 				page_array[seg] = p;
 
 				if (!c_tx->use_sendpage) {
-					iov[seg].iov_base = kmap(p) + fp_off;
-					iov[seg].iov_len = plen;
+					void *kaddr = kmap_local_page(page_array[seg]);
 
 					/* Remember for later kunmap() */
 					kmap_mask |= BIT(seg);
+					iov[seg].iov_base = kaddr + fp_off;
+					iov[seg].iov_len = plen;
 
 					if (do_crc)
 						crypto_shash_update(
@@ -518,7 +526,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 							iov[seg].iov_base,
 							plen);
 				} else if (do_crc) {
-					kaddr = kmap_local_page(p);
+					kaddr = kmap_local_page(page_array[seg]);
 					crypto_shash_update(c_tx->mpa_crc_hd,
 							    kaddr + fp_off,
 							    plen);
@@ -542,7 +550,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 
 			if (++seg > (int)MAX_ARRAY) {
 				siw_dbg_qp(tx_qp(c_tx), "to many fragments\n");
-				siw_unmap_pages(page_array, kmap_mask);
+				siw_unmap_pages(iov, kmap_mask, MAX_ARRAY);
 				wqe->processed -= c_tx->bytes_unsent;
 				rv = -EMSGSIZE;
 				goto done_crc;
@@ -593,7 +601,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 	} else {
 		rv = kernel_sendmsg(s, &msg, iov, seg + 1,
 				    hdr_len + data_len + trl_len);
-		siw_unmap_pages(page_array, kmap_mask);
+		siw_unmap_pages(iov, kmap_mask, MAX_ARRAY);
 	}
 	if (rv < (int)hdr_len) {
 		/* Not even complete hdr pushed or negative rv */
Signature · 1 line
2.28.0.rc0.12.gb6a658bd00c9
#13 lore raw

Re: [PATCH V2] RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

[email protected] wrote: -----

Quoted 114 lines

To: "Jason Gunthorpe" <[email protected]>
From: [email protected]
Date: 06/22/2021 10:35PM
Cc: "Ira Weiny" <[email protected]>, "Mike Marciniszyn" <[email protected]>, "Dennis Dalessandro" <[email protected]>, "Doug Ledford" <[email protected]>, "Faisal Latif" <[email protected]>, "Shiraz Saleem" <[email protected]>, "Bernard Metzler" <[email protected]>, "Kamal Heib" <[email protected]>, [email protected], [email protected]
Subject: [EXTERNAL] [PATCH V2] RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS
protection is introduced.[1]

The use of kmap() in siw_tx_hdt() is all thread local therefore kmap_local_page() is a sufficient replacement and will work with pgmap
protected pages when those are implemented.

siw_tx_hdt() tracks pages used in a page_array. It uses that array to
unmap pages which were mapped on function exit. Not all entries in the
array are mapped and this is tracked in kmap_mask.

kunmap_local() takes a mapped address rather than a page. Alter siw_unmap_pages() to take the iov array to reuse the iov_base address of
each mapping. Use PAGE_MASK to get the proper address for kunmap_local().

kmap_local_page() mappings are tracked in a stack and must be unmapped
in the opposite order they were mapped in. Because segments are mapped
into the page array in increasing index order, modify siw_unmap_pages()
to unmap pages in decreasing order.

Use kmap_local_page() instead of kmap() to map pages in the page_array.

[1]
INVALID URI REMOVED lkml_20201009195033.3208459-2D59-2Dira.weiny-40intel.com_&d=DwIDAg&c= jf_iaSHvJObTbx-siA1ZOg&r=2TaYXQ0T-r8ZO1PP1alNwU_QJcRRLfmYTAgd3QCvqSc& m=ujJBVqPLdVdVxXvOu_PlFL3NVC0Znds3FgxyrtWJtwM&s=WZIBAdwlCqPIRjsNOGlly gQ6Hsug6ObgrWgO_nvBGyc&e=

---
Changes for V2:

	From Bernard
		Reuse iov[].iov_base rather than declaring another array of
		pointers and preserve the use of kmap_mask to know which iov's
		were kmapped.

---
drivers/infiniband/sw/siw/siw_qp_tx.c | 32
+++++++++++++++++----------

1 file changed, 20 insertions(+), 12 deletions(-)
drivers/infiniband/sw/siw/siw_qp_tx.c +13 -6
index db68a10d12cd..fd3b9e6a67d7 100644
@@ -396,13 +396,20 @@ static int siw_0copy_tx(struct socket *s,
 struct page **page,
 
 #define MAX_TRAILER (MPA_CRC_SIZE + 4)
 
-static void siw_unmap_pages(struct page **pp, unsigned long
 kmap_mask)
+static void siw_unmap_pages(struct kvec *iov, unsigned long
 kmap_mask, int len)
 {
-	while (kmap_mask) {
-		if (kmap_mask & BIT(0))
-			kunmap(*pp);
-		pp++;
-		kmap_mask >>= 1;
+	int i;
+
+	/*
+	 * Work backwards through the array to honor the kmap_local_page()
+	 * ordering requirements.
+	 */
+	for (i = (len-1); i >= 0; i--) {
+		if (kmap_mask & BIT(i)) {
+			unsigned long addr = (unsigned long)iov[i].iov_base;
+
+			kunmap_local((void *)(addr & PAGE_MASK));
+		}
	}
}
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -498,7 +505,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 					p = siw_get_upage(mem->umem,
 							  sge->laddr + sge_off);
 				if (unlikely(!p)) {
-					siw_unmap_pages(page_array, kmap_mask);
+					siw_unmap_pages(iov, kmap_mask, MAX_ARRAY);
 					wqe->processed -= c_tx->bytes_unsent;
 					rv = -EFAULT;
					goto done_crc;

drivers/infiniband/sw/siw/siw_qp_tx.c static int siw_tx_hdt(struct siw_iwarp_tx · line 517

Quoted 7 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -2
@@ -506,11 +513,12 @@ static int siw_tx_hdt(struct siw_iwarp_tx
 *c_tx, struct socket *s)
 				page_array[seg] = p;
 
 				if (!c_tx->use_sendpage) {
-					iov[seg].iov_base = kmap(p) + fp_off;
-					iov[seg].iov_len = plen;
+					void *kaddr = kmap_local_page(page_array[seg]);

we can use 'kmap_local_page(p)' here

drivers/infiniband/sw/siw/siw_qp_tx.c static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, · line 528

Quoted 12 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +2 -0
 					/* Remember for later kunmap() */
 					kmap_mask |= BIT(seg);
+					iov[seg].iov_base = kaddr + fp_off;
+					iov[seg].iov_len = plen;
 
 					if (do_crc)
 						crypto_shash_update(
@@ -518,7 +526,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 							iov[seg].iov_base,
 							plen);

This patch does not apply for me. Would I have to install first your [Patch 3/4] -- since the current patch references kmap_local_page() already? Maybe it is better to apply if it would be just one siw related patch in that series?

drivers/infiniband/sw/siw/siw_qp_tx.c static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s) · line 529

Quoted 3 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
 				} else if (do_crc) {
-					kaddr = kmap_local_page(p);
+					kaddr = kmap_local_page(page_array[seg]);

using 'kmap_local_page(p)' as you had it is straightforward and I would prefer it.

drivers/infiniband/sw/siw/siw_qp_tx.c static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, · line 554

Quoted 10 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +0 -0
 					crypto_shash_update(c_tx->mpa_crc_hd,
 							    kaddr + fp_off,
							    plen);
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -542,7 +550,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 
 			if (++seg > (int)MAX_ARRAY) {
 				siw_dbg_qp(tx_qp(c_tx), "to many fragments\n");
-				siw_unmap_pages(page_array, kmap_mask);
+				siw_unmap_pages(iov, kmap_mask, MAX_ARRAY);

to minimize the iterations over the byte array in 'siw_unmap_pages()', we may pass seg-1 instead of MAX_ARRAY

drivers/infiniband/sw/siw/siw_qp_tx.c static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, · line 605

Quoted 10 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +0 -0
 				wqe->processed -= c_tx->bytes_unsent;
 				rv = -EMSGSIZE;
				goto done_crc;
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -593,7 +601,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 	} else {
 		rv = kernel_sendmsg(s, &msg, iov, seg + 1,
 				    hdr_len + data_len + trl_len);
-		siw_unmap_pages(page_array, kmap_mask);
+		siw_unmap_pages(iov, kmap_mask, MAX_ARRAY);

to minimize the iterations over the byte array in 'siw_unmap_pages()', we may pass seg instead of MAX_ARRAY

Quoted 7 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +0 -0
 	}
 	if (rv < (int)hdr_len) {
		/* Not even complete hdr pushed or negative rv */
Signature · 1 line
2.28.0.rc0.12.gb6a658bd00c9
author
#14 lore raw

Re: [PATCH V2] RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

Quoted 6 lines · Bernard Metzler

[email protected] wrote: -----

Quoted 4 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +0 -0
@@ -506,11 +513,12 @@ static int siw_tx_hdt(struct siw_iwarp_tx
 *c_tx, struct socket *s)
 				page_array[seg] = p;
 
Quoted 6 lines

drivers/infiniband/sw/siw/siw_qp_tx.c static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s) · line 516

Quoted 4 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -2
 				if (!c_tx->use_sendpage) {
-					iov[seg].iov_base = kmap(p) + fp_off;
-					iov[seg].iov_len = plen;
+					void *kaddr = kmap_local_page(page_array[seg]);

we can use 'kmap_local_page(p)' here

Yes but I actually did this on purpose as it makes the code read clearly that the mapping is 'seg' element of the array. Do you prefer 'p' because this is a performant path?

Quoted 6 lines
Quoted 6 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +2 -0
 					/* Remember for later kunmap() */
 					kmap_mask |= BIT(seg);
+					iov[seg].iov_base = kaddr + fp_off;
+					iov[seg].iov_len = plen;
 
Quoted 11 lines

drivers/infiniband/sw/siw/siw_qp_tx.c static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, · line 528

Quoted 6 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +0 -0
 					if (do_crc)
 						crypto_shash_update(
@@ -518,7 +526,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 							iov[seg].iov_base,
 							plen);

This patch does not apply for me. Would I have to install first your [Patch 3/4] -- since the current patch references kmap_local_page() already? Maybe it is better to apply if it would be just one siw related patch in that series?

Yes the other patch goes first. I split it out to make this more difficult change more reviewable. I could squash them as it is probably straight forward enough but I've been careful with this in other subsystems.

Jason, do you have any issue with squashing the 2 patches?

Quoted 9 lines

drivers/infiniband/sw/siw/siw_qp_tx.c static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s) · line 529

Quoted 6 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
 				} else if (do_crc) {
-					kaddr = kmap_local_page(p);
+					kaddr = kmap_local_page(page_array[seg]);

using 'kmap_local_page(p)' as you had it is straightforward and I would prefer it.

OK. I think this reads cleaner but I can see 'p' being more performant.

Quoted 7 lines
Quoted 7 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +0 -0
 					crypto_shash_update(c_tx->mpa_crc_hd,
 							    kaddr + fp_off,
							    plen);
(quoted diff) +0 -0
@@ -542,7 +550,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 
Quoted 7 lines

drivers/infiniband/sw/siw/siw_qp_tx.c static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s) · line 553

Quoted 4 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
 			if (++seg > (int)MAX_ARRAY) {
 				siw_dbg_qp(tx_qp(c_tx), "to many fragments\n");
-				siw_unmap_pages(page_array, kmap_mask);
+				siw_unmap_pages(iov, kmap_mask, MAX_ARRAY);

to minimize the iterations over the byte array in 'siw_unmap_pages()', we may pass seg-1 instead of MAX_ARRAY

Sounds good.

Quoted 15 lines

drivers/infiniband/sw/siw/siw_qp_tx.c static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, · line 605

Quoted 12 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +0 -0
 				wqe->processed -= c_tx->bytes_unsent;
 				rv = -EMSGSIZE;
				goto done_crc;
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -593,7 +601,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 	} else {
 		rv = kernel_sendmsg(s, &msg, iov, seg + 1,
 				    hdr_len + data_len + trl_len);
-		siw_unmap_pages(page_array, kmap_mask);
+		siw_unmap_pages(iov, kmap_mask, MAX_ARRAY);

to minimize the iterations over the byte array in 'siw_unmap_pages()', we may pass seg instead of MAX_ARRAY

Will do.

Thanks for the review! :-D
Ira

#15 lore raw

PATCH v3 1/1 RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

In reply to [email protected]

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS protection is introduced.[1]

The use of kmap() in siw_tx_hdt() is all thread local therefore kmap_local_page() is a sufficient replacement and will work with pgmap protected pages when those are implemented.

siw_tx_hdt() tracks pages used in a page_array. It uses that array to unmap pages which were mapped on function exit. Not all entries in the array are mapped and this is tracked in kmap_mask.

kunmap_local() takes a mapped address rather than a page. Alter siw_unmap_pages() to take the iov array to reuse the iov_base address of each mapping. Use PAGE_MASK to get the proper address for kunmap_local().

kmap_local_page() mappings are tracked in a stack and must be unmapped in the opposite order they were mapped in. Because segments are mapped into the page array in increasing index order, modify siw_unmap_pages() to unmap pages in decreasing order.

Use kmap_local_page() instead of kmap() to map pages in the page_array.

[1] https://lore.kernel.org/lkml/[email protected]/

---
Jason, I went ahead and left this a separate patch. Let me know if you really want this and the other siw squashed.

Changes for V3:
From Bernard
Use 'p' in kmap_local_page() Use seg as length to siw_unmap_pages()

Changes for V2:
From Bernard
Reuse iov[].iov_base rather than declaring another array of pointers and preserve the use of kmap_mask to know which iov's were kmapped.

---
 drivers/infiniband/sw/siw/siw_qp_tx.c | 30 +++++++++++++++++----------
 1 file changed, 19 insertions(+), 11 deletions(-)
drivers/infiniband/sw/siw/siw_qp_tx.c +19 -11
index db68a10d12cd..89a5b75f7254 100644
@@ -396,13 +396,20 @@ static int siw_0copy_tx(struct socket *s, struct page **page,
 
 #define MAX_TRAILER (MPA_CRC_SIZE + 4)
 
-static void siw_unmap_pages(struct page **pp, unsigned long kmap_mask)
+static void siw_unmap_pages(struct kvec *iov, unsigned long kmap_mask, int len)
 {
-	while (kmap_mask) {
-		if (kmap_mask & BIT(0))
-			kunmap(*pp);
-		pp++;
-		kmap_mask >>= 1;
+	int i;
+
+	/*
+	 * Work backwards through the array to honor the kmap_local_page()
+	 * ordering requirements.
+	 */
+	for (i = (len-1); i >= 0; i--) {
+		if (kmap_mask & BIT(i)) {
+			unsigned long addr = (unsigned long)iov[i].iov_base;
+
+			kunmap_local((void *)(addr & PAGE_MASK));
+		}
 	}
 }
 
@@ -498,7 +505,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 					p = siw_get_upage(mem->umem,
 							  sge->laddr + sge_off);
 				if (unlikely(!p)) {
-					siw_unmap_pages(page_array, kmap_mask);
+					siw_unmap_pages(iov, kmap_mask, seg);
 					wqe->processed -= c_tx->bytes_unsent;
 					rv = -EFAULT;
 					goto done_crc;
@@ -506,11 +513,12 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 				page_array[seg] = p;
 
 				if (!c_tx->use_sendpage) {
-					iov[seg].iov_base = kmap(p) + fp_off;
-					iov[seg].iov_len = plen;
+					void *kaddr = kmap_local_page(p);
 
 					/* Remember for later kunmap() */
 					kmap_mask |= BIT(seg);
+					iov[seg].iov_base = kaddr + fp_off;
+					iov[seg].iov_len = plen;
 
 					if (do_crc)
 						crypto_shash_update(
@@ -542,7 +550,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 
 			if (++seg > (int)MAX_ARRAY) {
 				siw_dbg_qp(tx_qp(c_tx), "to many fragments\n");
-				siw_unmap_pages(page_array, kmap_mask);
+				siw_unmap_pages(iov, kmap_mask, seg-1);
 				wqe->processed -= c_tx->bytes_unsent;
 				rv = -EMSGSIZE;
 				goto done_crc;
@@ -593,7 +601,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 	} else {
 		rv = kernel_sendmsg(s, &msg, iov, seg + 1,
 				    hdr_len + data_len + trl_len);
-		siw_unmap_pages(page_array, kmap_mask);
+		siw_unmap_pages(iov, kmap_mask, seg+1);
 	}
 	if (rv < (int)hdr_len) {
 		/* Not even complete hdr pushed or negative rv */
Signature · 1 line
2.28.0.rc0.12.gb6a658bd00c9
#16 lore raw

Re: [PATCH V3] RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

[email protected] wrote: -----

Quoted 148 lines

To: "Jason Gunthorpe" <[email protected]>
From: [email protected]
Date: 06/24/2021 12:16AM
Cc: "Ira Weiny" <[email protected]>, "Mike Marciniszyn" <[email protected]>, "Dennis Dalessandro" <[email protected]>, "Doug Ledford" <[email protected]>, "Faisal Latif" <[email protected]>, "Shiraz Saleem" <[email protected]>, "Bernard Metzler" <[email protected]>, "Kamal Heib" <[email protected]>, [email protected], [email protected]
Subject: [EXTERNAL] [PATCH V3] RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS
protection is introduced.[1]

The use of kmap() in siw_tx_hdt() is all thread local therefore kmap_local_page() is a sufficient replacement and will work with pgmap
protected pages when those are implemented.

siw_tx_hdt() tracks pages used in a page_array. It uses that array to
unmap pages which were mapped on function exit. Not all entries in the
array are mapped and this is tracked in kmap_mask.

kunmap_local() takes a mapped address rather than a page. Alter siw_unmap_pages() to take the iov array to reuse the iov_base address of
each mapping. Use PAGE_MASK to get the proper address for kunmap_local().

kmap_local_page() mappings are tracked in a stack and must be unmapped
in the opposite order they were mapped in. Because segments are mapped
into the page array in increasing index order, modify siw_unmap_pages()
to unmap pages in decreasing order.

Use kmap_local_page() instead of kmap() to map pages in the page_array.

[1]
INVALID URI REMOVED lkml_20201009195033.3208459-2D59-2Dira.weiny-40intel.com_&d=DwIDAg&c= jf_iaSHvJObTbx-siA1ZOg&r=2TaYXQ0T-r8ZO1PP1alNwU_QJcRRLfmYTAgd3QCvqSc& m=eI4Db7iSlEKRl4l5pYKwY5rL5WXWWxahhxNciwy2lRA&s=vo11VhOvYbAkABdhV6htX TmXgFZeWbBZAFnPDvg7Bzs&e=

---
Jason, I went ahead and left this a separate patch. Let me know if you really
want this and the other siw squashed.

Changes for V3:
From Bernard
Use 'p' in kmap_local_page() Use seg as length to siw_unmap_pages()

Changes for V2:
From Bernard
Reuse iov[].iov_base rather than declaring another array of pointers and preserve the use of kmap_mask to know which iov's were kmapped.
---
drivers/infiniband/sw/siw/siw_qp_tx.c | 30
+++++++++++++++++----------

1 file changed, 19 insertions(+), 11 deletions(-)
drivers/infiniband/sw/siw/siw_qp_tx.c +13 -6
index db68a10d12cd..89a5b75f7254 100644
@@ -396,13 +396,20 @@ static int siw_0copy_tx(struct socket *s,
 struct page **page,
 
 #define MAX_TRAILER (MPA_CRC_SIZE + 4)
 
-static void siw_unmap_pages(struct page **pp, unsigned long
 kmap_mask)
+static void siw_unmap_pages(struct kvec *iov, unsigned long
 kmap_mask, int len)
 {
-	while (kmap_mask) {
-		if (kmap_mask & BIT(0))
-			kunmap(*pp);
-		pp++;
-		kmap_mask >>= 1;
+	int i;
+
+	/*
+	 * Work backwards through the array to honor the kmap_local_page()
+	 * ordering requirements.
+	 */
+	for (i = (len-1); i >= 0; i--) {
+		if (kmap_mask & BIT(i)) {
+			unsigned long addr = (unsigned long)iov[i].iov_base;
+
+			kunmap_local((void *)(addr & PAGE_MASK));
+		}
	}
}
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -498,7 +505,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 					p = siw_get_upage(mem->umem,
 							  sge->laddr + sge_off);
 				if (unlikely(!p)) {
-					siw_unmap_pages(page_array, kmap_mask);
+					siw_unmap_pages(iov, kmap_mask, seg);
 					wqe->processed -= c_tx->bytes_unsent;
 					rv = -EFAULT;
					goto done_crc;
drivers/infiniband/sw/siw/siw_qp_tx.c +3 -2
@@ -506,11 +513,12 @@ static int siw_tx_hdt(struct siw_iwarp_tx
 *c_tx, struct socket *s)
 				page_array[seg] = p;
 
 				if (!c_tx->use_sendpage) {
-					iov[seg].iov_base = kmap(p) + fp_off;
-					iov[seg].iov_len = plen;
+					void *kaddr = kmap_local_page(p);
 
 					/* Remember for later kunmap() */
 					kmap_mask |= BIT(seg);
+					iov[seg].iov_base = kaddr + fp_off;
+					iov[seg].iov_len = plen;
 
 					if (do_crc)
						crypto_shash_update(
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -542,7 +550,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 
 			if (++seg > (int)MAX_ARRAY) {
 				siw_dbg_qp(tx_qp(c_tx), "to many fragments\n");
-				siw_unmap_pages(page_array, kmap_mask);
+				siw_unmap_pages(iov, kmap_mask, seg-1);
 				wqe->processed -= c_tx->bytes_unsent;
 				rv = -EMSGSIZE;
				goto done_crc;

drivers/infiniband/sw/siw/siw_qp_tx.c static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, · line 605

Quoted 6 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -593,7 +601,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 	} else {
 		rv = kernel_sendmsg(s, &msg, iov, seg + 1,
 				    hdr_len + data_len + trl_len);
-		siw_unmap_pages(page_array, kmap_mask);
+		siw_unmap_pages(iov, kmap_mask, seg+1);

seg+1 is one to many, since the last segment references the iWarp trailer (CRC). There are 2 reason for this multi-segment processing in the transmit path. (1) efficiency and (2) MTU based packet framing. The iov contains the complete iWarp frame with header, (potentially multiple) data fragments, and the CRC. It gets pushed to TCP in one go, praying for iWarp framing stays intact (which most time works). So the code can collect data form multiple SGE's of a WRITE or SEND and tries putting those into one frame, if MTU allows, and adds header and trailer.
The last segment (seg + 1) references the CRC, which is never kmap'ed.

I'll try the code next days, but it looks good otherwise!

Thanks very much!

Quoted 7 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +0 -0
 	}
 	if (rv < (int)hdr_len) {
		/* Not even complete hdr pushed or negative rv */
Signature · 1 line
2.28.0.rc0.12.gb6a658bd00c9
author
#17 lore raw

Re: [PATCH V3] RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

Quoted 19 lines · Bernard Metzler

drivers/infiniband/sw/siw/siw_qp_tx.c static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, · line 605

Quoted 8 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -593,7 +601,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 	} else {
 		rv = kernel_sendmsg(s, &msg, iov, seg + 1,
 				    hdr_len + data_len + trl_len);
-		siw_unmap_pages(page_array, kmap_mask);
+		siw_unmap_pages(iov, kmap_mask, seg+1);

seg+1 is one to many, since the last segment references the iWarp trailer (CRC). There are 2 reason for this multi-segment processing in the transmit path. (1) efficiency and (2) MTU based packet framing. The iov contains the complete iWarp frame with header, (potentially multiple) data fragments, and the CRC. It gets pushed to TCP in one go, praying for iWarp framing stays intact (which most time works). So the code can collect data form multiple SGE's of a WRITE or SEND and tries putting those into one frame, if MTU allows, and adds header and trailer.

Quoted 1 line

The last segment (seg + 1) references the CRC, which is never kmap'ed.

siw_unmap_pages() take a length and seg is the index...

But ok so a further optimization...

Fair enough.

Quoted 2 lines

I'll try the code next days, but it looks good otherwise!

I believe this will work though.

Ira

Quoted 8 lines

Thanks very much!

Quoted 7 lines
drivers/infiniband/sw/siw/siw_qp_tx.c +0 -0
 	}
 	if (rv < (int)hdr_len) {
		/* Not even complete hdr pushed or negative rv */
Signature · 1 line
2.28.0.rc0.12.gb6a658bd00c9
#18 lore raw

PATCH v4 1/1 RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

In reply to [email protected]

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS protection is introduced.[1]

The use of kmap() in siw_tx_hdt() is all thread local therefore kmap_local_page() is a sufficient replacement and will work with pgmap protected pages when those are implemented.

siw_tx_hdt() tracks pages used in a page_array. It uses that array to unmap pages which were mapped on function exit. Not all entries in the array are mapped and this is tracked in kmap_mask.

kunmap_local() takes a mapped address rather than a page. Alter siw_unmap_pages() to take the iov array to reuse the iov_base address of each mapping. Use PAGE_MASK to get the proper address for kunmap_local().

kmap_local_page() mappings are tracked in a stack and must be unmapped in the opposite order they were mapped in. Because segments are mapped into the page array in increasing index order, modify siw_unmap_pages() to unmap pages in decreasing order.

Use kmap_local_page() instead of kmap() to map pages in the page_array.

[1] https://lore.kernel.org/lkml/[email protected]/

---
Changes for V4:
From Bernard
Further optimize siw_unmap_pages() by eliminating the CRC page from the iov array.

Changes for V3:
From Bernard
Use 'p' in kmap_local_page() Use seg as length to siw_unmap_pages()

Changes for V2:
From Bernard
Reuse iov[].iov_base rather than declaring another array of pointers and preserve the use of kmap_mask to know which iov's were kmapped.

---
 drivers/infiniband/sw/siw/siw_qp_tx.c | 30 +++++++++++++++++----------
 1 file changed, 19 insertions(+), 11 deletions(-)
drivers/infiniband/sw/siw/siw_qp_tx.c +19 -11
index db68a10d12cd..1f4e60257700 100644
@@ -396,13 +396,20 @@ static int siw_0copy_tx(struct socket *s, struct page **page,
 
 #define MAX_TRAILER (MPA_CRC_SIZE + 4)
 
-static void siw_unmap_pages(struct page **pp, unsigned long kmap_mask)
+static void siw_unmap_pages(struct kvec *iov, unsigned long kmap_mask, int len)
 {
-	while (kmap_mask) {
-		if (kmap_mask & BIT(0))
-			kunmap(*pp);
-		pp++;
-		kmap_mask >>= 1;
+	int i;
+
+	/*
+	 * Work backwards through the array to honor the kmap_local_page()
+	 * ordering requirements.
+	 */
+	for (i = (len-1); i >= 0; i--) {
+		if (kmap_mask & BIT(i)) {
+			unsigned long addr = (unsigned long)iov[i].iov_base;
+
+			kunmap_local((void *)(addr & PAGE_MASK));
+		}
 	}
 }
 
@@ -498,7 +505,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 					p = siw_get_upage(mem->umem,
 							  sge->laddr + sge_off);
 				if (unlikely(!p)) {
-					siw_unmap_pages(page_array, kmap_mask);
+					siw_unmap_pages(iov, kmap_mask, seg);
 					wqe->processed -= c_tx->bytes_unsent;
 					rv = -EFAULT;
 					goto done_crc;
@@ -506,11 +513,12 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 				page_array[seg] = p;
 
 				if (!c_tx->use_sendpage) {
-					iov[seg].iov_base = kmap(p) + fp_off;
-					iov[seg].iov_len = plen;
+					void *kaddr = kmap_local_page(p);
 
 					/* Remember for later kunmap() */
 					kmap_mask |= BIT(seg);
+					iov[seg].iov_base = kaddr + fp_off;
+					iov[seg].iov_len = plen;
 
 					if (do_crc)
 						crypto_shash_update(
@@ -542,7 +550,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 
 			if (++seg > (int)MAX_ARRAY) {
 				siw_dbg_qp(tx_qp(c_tx), "to many fragments\n");
-				siw_unmap_pages(page_array, kmap_mask);
+				siw_unmap_pages(iov, kmap_mask, seg-1);
 				wqe->processed -= c_tx->bytes_unsent;
 				rv = -EMSGSIZE;
 				goto done_crc;
@@ -593,7 +601,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
 	} else {
 		rv = kernel_sendmsg(s, &msg, iov, seg + 1,
 				    hdr_len + data_len + trl_len);
-		siw_unmap_pages(page_array, kmap_mask);
+		siw_unmap_pages(iov, kmap_mask, seg);
 	}
 	if (rv < (int)hdr_len) {
 		/* Not even complete hdr pushed or negative rv */
Signature · 1 line
2.28.0.rc0.12.gb6a658bd00c9
Reviewed-by on v4 1/1
#19 lore raw

Re: [PATCH V4] RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

[email protected] wrote: -----

Quoted 157 lines

To: "Jason Gunthorpe" <[email protected]>
From: [email protected]
Date: 06/24/2021 07:48PM
Cc: "Ira Weiny" <[email protected]>, "Mike Marciniszyn" <[email protected]>, "Dennis Dalessandro" <[email protected]>, "Doug Ledford" <[email protected]>, "Faisal Latif" <[email protected]>, "Shiraz Saleem" <[email protected]>, "Bernard Metzler" <[email protected]>, "Kamal Heib" <[email protected]>, [email protected], [email protected]
Subject: [EXTERNAL] [PATCH V4] RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS
protection is introduced.[1]

The use of kmap() in siw_tx_hdt() is all thread local therefore kmap_local_page() is a sufficient replacement and will work with pgmap
protected pages when those are implemented.

siw_tx_hdt() tracks pages used in a page_array. It uses that array to
unmap pages which were mapped on function exit. Not all entries in the
array are mapped and this is tracked in kmap_mask.

kunmap_local() takes a mapped address rather than a page. Alter siw_unmap_pages() to take the iov array to reuse the iov_base address of
each mapping. Use PAGE_MASK to get the proper address for kunmap_local().

kmap_local_page() mappings are tracked in a stack and must be unmapped
in the opposite order they were mapped in. Because segments are mapped
into the page array in increasing index order, modify siw_unmap_pages()
to unmap pages in decreasing order.

Use kmap_local_page() instead of kmap() to map pages in the page_array.

[1]
INVALID URI REMOVED lkml_20201009195033.3208459-2D59-2Dira.weiny-40intel.com_&d=DwIDAg&c= jf_iaSHvJObTbx-siA1ZOg&r=2TaYXQ0T-r8ZO1PP1alNwU_QJcRRLfmYTAgd3QCvqSc& m=01QnZvj05j7vvgDChewVpHJlDytiIFuttai7VRUdJMs&s=zS4nDlvF_3MDi9wu7GaL6 qooDhiboqP5ii5ozBeDpLE&e=

---
Changes for V4:
From Bernard
Further optimize siw_unmap_pages() by eliminating the CRC page from the iov array.

Changes for V3:
From Bernard
Use 'p' in kmap_local_page() Use seg as length to siw_unmap_pages()

Changes for V2:
From Bernard
Reuse iov[].iov_base rather than declaring another array of pointers and preserve the use of kmap_mask to know which iov's were kmapped.
---
drivers/infiniband/sw/siw/siw_qp_tx.c | 30
+++++++++++++++++----------

1 file changed, 19 insertions(+), 11 deletions(-)
drivers/infiniband/sw/siw/siw_qp_tx.c +13 -6
index db68a10d12cd..1f4e60257700 100644
@@ -396,13 +396,20 @@ static int siw_0copy_tx(struct socket *s,
 struct page **page,
 
 #define MAX_TRAILER (MPA_CRC_SIZE + 4)
 
-static void siw_unmap_pages(struct page **pp, unsigned long
 kmap_mask)
+static void siw_unmap_pages(struct kvec *iov, unsigned long
 kmap_mask, int len)
 {
-	while (kmap_mask) {
-		if (kmap_mask & BIT(0))
-			kunmap(*pp);
-		pp++;
-		kmap_mask >>= 1;
+	int i;
+
+	/*
+	 * Work backwards through the array to honor the kmap_local_page()
+	 * ordering requirements.
+	 */
+	for (i = (len-1); i >= 0; i--) {
+		if (kmap_mask & BIT(i)) {
+			unsigned long addr = (unsigned long)iov[i].iov_base;
+
+			kunmap_local((void *)(addr & PAGE_MASK));
+		}
	}
}
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -498,7 +505,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 					p = siw_get_upage(mem->umem,
 							  sge->laddr + sge_off);
 				if (unlikely(!p)) {
-					siw_unmap_pages(page_array, kmap_mask);
+					siw_unmap_pages(iov, kmap_mask, seg);
 					wqe->processed -= c_tx->bytes_unsent;
 					rv = -EFAULT;
					goto done_crc;
drivers/infiniband/sw/siw/siw_qp_tx.c +3 -2
@@ -506,11 +513,12 @@ static int siw_tx_hdt(struct siw_iwarp_tx
 *c_tx, struct socket *s)
 				page_array[seg] = p;
 
 				if (!c_tx->use_sendpage) {
-					iov[seg].iov_base = kmap(p) + fp_off;
-					iov[seg].iov_len = plen;
+					void *kaddr = kmap_local_page(p);
 
 					/* Remember for later kunmap() */
 					kmap_mask |= BIT(seg);
+					iov[seg].iov_base = kaddr + fp_off;
+					iov[seg].iov_len = plen;
 
 					if (do_crc)
						crypto_shash_update(
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -542,7 +550,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 
 			if (++seg > (int)MAX_ARRAY) {
 				siw_dbg_qp(tx_qp(c_tx), "to many fragments\n");
-				siw_unmap_pages(page_array, kmap_mask);
+				siw_unmap_pages(iov, kmap_mask, seg-1);
 				wqe->processed -= c_tx->bytes_unsent;
 				rv = -EMSGSIZE;
				goto done_crc;
drivers/infiniband/sw/siw/siw_qp_tx.c +1 -1
@@ -593,7 +601,7 @@ static int siw_tx_hdt(struct siw_iwarp_tx *c_tx,
 struct socket *s)
 	} else {
 		rv = kernel_sendmsg(s, &msg, iov, seg + 1,
 				    hdr_len + data_len + trl_len);
-		siw_unmap_pages(page_array, kmap_mask);
+		siw_unmap_pages(iov, kmap_mask, seg);
 	}
 	if (rv < (int)hdr_len) {
Quoted 5 lines
		/* Not even complete hdr pushed or negative rv */
Signature · 1 line
2.28.0.rc0.12.gb6a658bd00c9

Sry my misconfigured email attached some HTML crap. So I did not reach the list.

Tested V4 which works as intended. Thanks, Ira!

author
#20 lore raw

Re: [PATCH V4] RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

Quoted 3 lines
Quoted 1 line

Sry my misconfigured email attached some HTML crap. So I did not reach the list.

NP...

Quoted 4 lines

Tested V4 which works as intended. Thanks, Ira!

Thanks!
Ira

maintainer Applied
#21 lore raw

Re: [PATCH V4] RDMA/siw: Convert siw_tx_hdt() to kmap_local_page()

In reply to [email protected]
Quoted 40 lines · On Thu, Jun 24, 2021 at 10:48:14AM -0700, [email protected] wrote:

From: Ira Weiny <[email protected]>

kmap() is being deprecated and will break uses of device dax after PKS protection is introduced.[1]

The use of kmap() in siw_tx_hdt() is all thread local therefore kmap_local_page() is a sufficient replacement and will work with pgmap protected pages when those are implemented.

siw_tx_hdt() tracks pages used in a page_array. It uses that array to unmap pages which were mapped on function exit. Not all entries in the array are mapped and this is tracked in kmap_mask.

kunmap_local() takes a mapped address rather than a page. Alter siw_unmap_pages() to take the iov array to reuse the iov_base address of each mapping. Use PAGE_MASK to get the proper address for kunmap_local().

kmap_local_page() mappings are tracked in a stack and must be unmapped in the opposite order they were mapped in. Because segments are mapped into the page array in increasing index order, modify siw_unmap_pages() to unmap pages in decreasing order.

Use kmap_local_page() instead of kmap() to map pages in the page_array.

[1] https://lore.kernel.org/lkml/[email protected]/

---
Changes for V4:
From Bernard
Further optimize siw_unmap_pages() by eliminating the CRC page from the iov array.

Changes for V3:
From Bernard
Use 'p' in kmap_local_page() Use seg as length to siw_unmap_pages()

Quoted 8 lines

Changes for V2:
From Bernard
Reuse iov[].iov_base rather than declaring another array of pointers and preserve the use of kmap_mask to know which iov's were kmapped.

---
 drivers/infiniband/sw/siw/siw_qp_tx.c | 30 +++++++++++++++++----------
 1 file changed, 19 insertions(+), 11 deletions(-)

Applied to for-next, thanks

Jason